A Free AI Agent for Business Is Not the Risk. Where Your Rules Live Is

On 28 September Meta announced the Meta Enterprise Platform, which Mark Zuckerberg calls "the next major pillar of our business", and a day later it introduced Muse for Small Business. The question worth asking is not whether the agent is any good, but where your business rules end up living.

What a free AI agent for business actually costs you

An AI agent is only as safe as the place your business rules are stored. Meta says Muse for Small Business connects to tools owners already use, including Shopify, Stripe, Intuit QuickBooks, Slack and Notion, supports custom connectors, and is free for most of what people need, with paid plans above that. TechCrunch's view is that the platform could help Meta see a return on its AI spend. Free is a fair price for a good tool.

The cost arrives later. Every pricing exception, every approval habit and every supplier rule you teach an agent inside someone else's product becomes something you cannot easily take out. Leaving turns into a rebuild of your own know-how.

South African owners are already forming the habit. Xero's 2026 State of South African Small Business report, which surveyed more than 400 businesses in early 2026, found that "more than half (52%) are using AI tools daily or weekly", and 44% use AI to automate processes. Those figures are about AI tools, not agents. The habit is what matters. Agents are the next step, and the decisions made now will set what is expensive to change later.

The approval gate should be your design, not a setting

Meta says nothing publishes, sends or spends without your approval. That is a good feature. It is also a switch inside Meta's app, built the way Meta decided, and it does not travel with you if you move.

We proposed a different shape to a South African education app client. The proposal keeps the AI out of the live customer path. In the proposed flow, users are served questions from an approved database, so nobody waits on a model. Background generation is triggered by real usage, so cost follows demand instead of running ahead of it. A human admin approves each batch in a portal and can mark standout items as gold standard, and those approved examples are fed back in as prompt examples, with no costly retraining.

The point is what you own. Because the approved questions live in a database you own, the model behind them is easier to swap. The approved data and the approval gate are the asset. The AI model is a replaceable worker behind them. Swap the worker and the asset stays put.

That is the same reasoning behind avoiding AI vendor lock-in: the model should be the part you can change on a Tuesday afternoon.

How to set this up before you adopt any agent

Meta states that Muse is available in the US and Canada. Wherever you are, the architecture is worth deciding before convenience decides it for you.

Three rules cover most of it. Keep your customer, order and pricing data in systems you own. Keep the workflow logic, the rules that say who approves what and when, in your own software rather than in an agent's settings. And connect any agent, Meta's or another, through a narrow connection you can unplug, so replacing it is a change of one connector and not a change of how your business runs.

None of this needs a large build. It needs a system of record you control and an approval step you designed. If you have neither, that is a job for custom software built around your process, and it is worth doing before the agent, not after.

Use the free agent. Just make sure that on the day you stop using it, your business rules are still yours.